Privacy constraints block lots of valuable research
Many valuable research questions remain difficult to pursue because the required data cannot be shared freely. Patient records, genomes, financial transactions, and rare-disease registries contain sensitive information that institutions are responsible for protecting.
Many useful questions also span institutions. A rare-disease study may need cohorts from several hospitals. A fraud model may need patterns from more than one bank. A public-health estimate may depend on sources that cannot publish individual records. The analysis may be technically possible while privacy, security, and governance constraints prevent the underlying data from being pooled.
A different default: compute on data you can't see
A common approach is to move the data to a trusted environment. The Blind Machine offers another model: participants keep their source data and secret keys on their own machines while the hosted service works with ciphertext.
It is a platform for governed computation on encrypted data. Each participant processes and encrypts records locally. The hosted service receives ciphertext, performs the approved computation, and returns an encrypted result. Plaintext and secret keys remain with the participants.
How it works
Four properties define the model. Cryptography protects the confidentiality of participant data, while governance rules constrain what may run and what may be released. The current version still trusts the hosted service for the correctness of the aggregate value.
Your data stays yours. Participants generate and retain their secret keys, process their records locally, and send ciphertext to the service. The hosted workflow does not require participants to upload plaintext records or secret keys.
The server computes blind. Homomorphic encryption lets a computer do real arithmetic directly on encrypted numbers and return an encrypted answer — one that only your key can open. The service adds and combines your ciphertext without ever learning what it holds. A machine that works without sight.
Every result carries a receipt you can check. Each computation returns a tamper-evident certificate that binds four things together: the computation, the committed encrypted inputs, the encrypted result, and the governing release policy. The certificate can be verified offline on a participant's machine.
You read the code before it runs. Every computation is a signed, content-addressed "application" — a small, inspectable bundle that pins the code, its sealed environment, and its release policy into one digest. Participants can inspect the bundle and verify that its digest identifies the bytes approved for the run.
Governance comes before computation. Cohort commitments, minimum cohort sizes, run caps, and aggregate-only release rules are evaluated before a computation begins. A study cannot run until its configured governance conditions are satisfied.
Why it runs on a laptop
Homomorphic encryption can require substantial time and memory, especially when a workload is expressed as a deep general-purpose encrypted circuit.
The Blind Machine reduces that cost in two ways.
Push the work to the edges. Before anything gets encrypted, each participant does the heavy lifting over their own plaintext — the counting, the summarizing, the per-record math — while it's still cheap and fast, right where the data already lives. Only the small, already-reduced quantities get encrypted and sent up. This is federated computing: the expensive part happens on a laptop, in the clear, at home. The server only ever combines the tiny encrypted remainders.
Use the simplest encryption scheme sufficient for the workload. Many supported analyses require sums or multiplication by a public number rather than deep multiplicative circuits. Additive homomorphic encryption is less expressive than fully homomorphic encryption, but it can answer these questions with lower computational cost.
For the workloads demonstrated so far, combining local reduction with additive homomorphic encryption makes computation practical on a laptop, with measured operations ranging from seconds to minutes.
Feasibility demonstrated
The Blind Machine has reproduced two published, peer-reviewed encrypted-genomics methods on the platform without changing its architecture for either reproduction.
The first is an encrypted polygenic-risk-score method (HEPRS). The second is the encrypted genome-wide association study described by Blatt et al. HEPRS returned risk scores bit-for-bit identical to a plaintext reference. The association study reproduced the reference statistic exactly on its core test; its covariate-adjusted variant matched the reference to within five nines.
In the measured setup, the risk-score workload used roughly 150x less memory. For the association study, a linear extrapolation from a measured sub-second anchor projects about an hour on a single laptop core, compared with the 5.6 hours reported for the original method on a server node, or about two minutes across 31 SNP-block workers. This is a like-for-unlike projection — additive aggregation compared with a full encrypted circuit — rather than a measured head-to-head benchmark. Full numbers and methodology are in a forthcoming technical report.
These reproductions demonstrate feasibility for two genomics workloads. They do not establish the same performance or correctness for every application.
Who it's for
The first demonstrations focus on genomics, where useful analyses often depend on sensitive records held by multiple institutions: biobanks combining cohorts, hospitals running a study across sites, or rare-disease registries assembling enough participants for analysis.
The same model may apply when participants need to keep raw data private while releasing an approved aggregate:
- Healthcare — multi-hospital studies where patient records never leave the hospital that holds them.
- Finance — banks computing a shared fraud or risk signal without exposing a single customer ledger.
- Public statistics — a national figure aggregated from contributors, none of whom could publish their own rows.
- Collaborative machine learning — models informed by data that stays with its owner and never lands in a shared bucket.
Different fields, one shape: the raw data stays home, the approved answer comes out.
Open by design
The security-sensitive user surface is the open-source blind command-line tool, MIT-licensed at github.com/blindmachine/blind. It supports account creation, study setup, encrypted contribution, computation, and certificate verification from a participant's machine.
The website provides a browser view of studies, jobs, and results. Operations involving plaintext and secret keys remain in the open-source CLI, where participants can inspect the code that handles them.
Get started
Three ways to evaluate the system:
Run the demo. Install the CLI and execute the full workflow on a laptop — one researcher, two data providers, encryption, computation, and certificate verification. Compare the encrypted workflow's result with the included plaintext reference.
Read an application. Inspect the signed bundle before using it with your own ciphertext: code, sealed environment, release policy, and digest.
Build your own. Bring your question and write the small piece of math that answers it. The platform handles the encryption, the governance, and the certificate.
When a research question depends on data that cannot be shared directly, encrypted computation offers a way to collaborate while keeping the underlying records local. Bring the question. The data can stay home.